| # FastWAM Base | |
| This directory contains the complete FastWAM initialization used by the | |
| FluxVLA FastWAM LIBERO full-finetuning recipe. | |
| ## Contents | |
| - `fastwam_base_full.safetensors`: bf16 FastWAM state dict initialized with | |
| seed 42. It contains the video DiT, VAE, text encoder, and ActionDiT | |
| parameters. | |
| - `tokenizer/`: UMT5 tokenizer files used for online text encoding. | |
| - `manifest.json`: package metadata and integrity information. | |
| The package intentionally excludes task-specific text-embedding caches. Text | |
| prompts are encoded online by the model loaded from the complete checkpoint. | |
| After downloading this directory to `./checkpoints/fastwam_base`, configure | |
| FastWAM with: | |
| ```python | |
| pretrained_name_or_path = ( | |
| './checkpoints/fastwam_base/fastwam_base_full.safetensors') | |
| tokenizer_path = './checkpoints/fastwam_base/tokenizer' | |
| ``` | |
| The corresponding FluxVLA recipe is | |
| `configs/fastwam/fastwam_libero_full_finetune.py`. | |